Reading Research Papers

Read with a purpose: deciding whether a paper is relevant, understanding its result, or using its method. You do not need the same level of detail for each.

Read it three times

Keshav's three-pass method separates these levels:

Scan the abstract, introduction, headings, figures and conclusion. Identify the problem and main claim.
Read the argument and evaluation, leaving difficult proofs or implementation details for later. Check that you can explain the result.
Work through the details when you need to implement, extend or assess the work closely.

Stop when you have enough for your purpose.

What each section is for

Section	Look for
Abstract	Main claim
Introduction	Problem and contribution
Related work	Context and earlier approaches
Method	What the authors did
Evaluation	Evidence for the claim
Conclusion	Results and limitations

Reading the evaluation properly

Check the baselines, datasets, assumptions and measures. Look for variation across runs, missing comparisons and limitations. Read graph axes carefully; a reported improvement may not matter for your use case.

When the maths defeats you

Try a small concrete case, read the surrounding explanation, or find a textbook or survey introducing the notation. Mark details to revisit rather than stopping at every unfamiliar step.

Taking notes that survive

Record the problem, main idea, assumptions and relevance to your work. Use your own words and keep the citation with your notes.

Checking a claim

Distinguish the result the authors demonstrate from the broader implication they suggest. A benchmark on one workload does not establish performance on every workload. A theorem only applies when its assumptions hold.

For an empirical claim, identify what changed between the proposed method and the baseline. Check whether tuning effort, hardware and data access were comparable. An ablation removes part of a method to examine its contribution.

Reproducing the result

Look for code, data, configuration and version information. Record any steps the paper leaves unspecified. Reproducing a result and independently confirming its explanation are different goals.

Start with the smallest result that would show whether your setup is correct. Save commands and outputs so another person can repeat the run.

A short note format

Problem and setting.
Main claim and how it is supported.
Assumptions and limitations.
What differs from earlier work.
What you would need to use or test the idea.

Keep unresolved questions explicit. A clear question is more useful than a paragraph of copied terminology.

Finding papers worth reading

Google Scholar and Semantic Scholar help follow citations.
DBLP indexes computer science publications.
arXiv hosts preprints; check publication and review status.

A survey is a useful starting point in an unfamiliar field. Use Cambridge's library access for publications covered by University subscriptions.